Analysis of Thermal Side-Channel Attacks on Analog/Digital Computing-in-Memory Accelerators

Compute-in-memory (CIM) promises to tackle the energy-efficiency barriers encountered by Von-Neuman computing, especially for machine learning workloads. One option of implementing CIM works by storing neural network weights in emerging memristor devices, connected together to form crossbars. This structure can conduct matrix-vector multiplications (MVM) in-place. The calculation causes heat to be emitted from the memristors, leading to a thermal side-channel. An attacker could sense the temperatures of the memristors, allowing recovery of the stored weights. We develop a fast and accurate thermal simulator, and evaluate the effectiveness of potential thermal side-channels. We evaluate differential weight mapping with different quantization levels. Results include recovery of weights with 70-100% accuracy for neural network layers.